Microsoft AB-100 (Agentic AI Business Solutions Architect) Practice Test 83

Question 820

A test plan focuses on a previously missed failure: Overall accuracy conceals disproportionate error. It occurred when a classifier underperforms on one language group. Which check directly validates that behavior?

  1. Evaluate a representative multilingual holdout
  2. Test equivalent requests in represented languages
  3. Try injection, invalid inputs and out-of-scope actions
  4. Disable knowledge access and verify safe no-evidence behavior

Explanation

Correct Answer

A:

The scenario is that a classifier underperforms on one language group; the defect to rule out is that overall accuracy conceals disproportionate error; testing via evaluating a representative multilingual holdout challenges the design dependency where segmented evaluations reveal uneven model behavior.

Incorrect Answers

B:

The verification concerning an AI triage model under-serves one customer language checks segment-aware evaluations surface fairness risks against overall performance masks unequal escalation rates; its result says little about the observed issue where overall accuracy conceals disproportionate error.

C:

In a solution where benchmarks cover only cooperative user wording, trying injection, invalid inputs and out-of-scope actions could catch the defect that a malicious instruction triggers a privileged tool; the current release gate must challenge the failure where overall accuracy conceals disproportionate error.

D:

The suggested test serves policy answers must be tied to a current source, in which the agent confidently answers when retrieval fails; the present scenario instead concerns the reported problem where overall accuracy conceals disproportionate error.

Question 821

A business stakeholder describes this operating situation: Analysts compare unstructured reports but must retain human interpretation. Testing instead shows that narratives present speculation as if it were validated insight. What design approach is most appropriate?

  1. Use an analyst-facing grounded assistant for exploratory insights
  2. Use an autonomous agent only for bounded reversible tasks
  3. Use A2A for agent-to-agent collaboration across runtimes
  4. Separate retrieved data from trusted system instructions

Explanation

Correct Answer

A:

The defect is that narratives present speculation as if it were validated insight; the expected business case is that analysts compare unstructured reports but must retain human interpretation; the design fits because retrieval supplies supporting evidence while analysts own analytical decisions.

Incorrect Answers

B:

In the case where work can proceed from events without human initiation but needs bounded autonomy, the alternate design depends on event triggers and tool permissions can automate predefined safe steps within explicit limits; it does not correct the target failure, narratives present speculation as if it were validated insight.

C:

An agent interoperability protocol supports explicit task delegation between agents; its main purpose is independently hosted agents must exchange tasks and results, not correction of the reported defect where narratives present speculation as if it were validated insight in the target process.

D:

For third-party documents include sentences telling the model to ignore rules, this other approach relies on the fact that trust boundaries treat source text as evidence rather than authority; the issue here is narratives present speculation as if it were validated insight; that mechanism does not establish the required retrieval supplies supporting evidence while analysts own analytical decisions.

Question 822

The architecture board has evidence that average handling-time savings disappear during rework. The affected situation is: AI drafts requests quickly but specialists must correct some outputs. What change most directly establishes the missing behavior?

  1. Include human review and exception handling in unit economics
  2. Account for capacity, consumption and monitoring in TCO
  3. Retain process-specific human review alongside prebuilt automation
  4. Maintain system-of-record ownership across Dynamics integrations

Explanation

Correct Answer

A:

AI drafts requests quickly but specialists must correct some outputs; its failing behavior is that average handling-time savings disappear during rework; applying the underlying design principle—end-to-end unit cost includes review, escalation and retry effort—connects the requirement to an implementable control.

Incorrect Answers

B:

For an agent scales from a trial to enterprise usage, this other approach relies on the fact that workload forecasts include model usage, connectors, storage and observability; the issue here is average handling-time savings disappear during rework; that mechanism does not establish the required end-to-end unit cost includes review, escalation and retry effort.

C:

A review step preserves accountability without discarding assisted drafting; its main purpose is standard capabilities draft customer communications but policy requires approval, not correction of the reported defect where average handling-time savings disappear during rework in the target process.

D:

Parallel agent updates create conflicting authoritative values; the alternate design can address that situation because field ownership and integration contracts prevent competing writes, whereas this question needs a solution to average handling-time savings disappear during rework.

Question 823

Before introducing a fix, the team must explain why every agent reimplements the same proprietary tool signatures. The affected setting is: A Copilot Studio agent needs a reusable interface to governed operations. What should be inspected?

  1. Inspect table metadata and relationship mappings across applications
  2. Compare orchestration requirements with host extensibility limits
  3. Capture advertised capabilities and actual authorized tool calls
  4. Compare host limitations against the requirements matrix

Explanation

Correct Answer

C:

The reported fault is that every agent reimplements the same proprietary tool signatures in a system where a Copilot Studio agent needs a reusable interface to governed operations; evidence from capturing advertised capabilities and actual authorized tool calls can show whether MCP presents supported tools and resources through a common integration contract.

Incorrect Answers

A:

Different solutions apply incompatible status and customer fields; examining that other path does not determine whether MCP presents supported tools and resources through a common integration contract, which is what distinguishes the issue in the stated use case.

B:

That probe evaluates custom-engine architecture controls runtime, model calls and integrations in a workflow where the solution demands its own model selection and hosting policies; the reported failure concerns the path for which MCP presents supported tools and resources through a common integration contract instead.

D:

In the process where a workflow needs proprietary scheduling, model choice and runtime control, the diagnostic procedure could explain a predefined host cannot meet the required orchestration contract; this particular incident instead turns on whether MCP presents supported tools and resources through a common integration contract and why every agent reimplements the same proprietary tool signatures.

Question 824

Before introducing a fix, the team must explain why generator tuning fails because the required documents were never retrieved. The affected setting is: Stakeholders report hallucinations but cannot isolate search errors. What should be inspected?

  1. Inspect step run histories and conditional branch outcomes
  2. Trace each release to labels, parameters and dataset hash
  3. Record recall at K and citation-grounding errors separately
  4. Inspect transcript alignment and transcription errors

Explanation

Correct Answer

C:

In the setting where stakeholders report hallucinations but cannot isolate search errors, recording recall at K and citation-grounding errors separately can isolate the failure in which generator tuning fails because the required documents were never retrieved, because labeled queries and expected evidence distinguish retrieval from synthesis.

Incorrect Answers

A:

An unconstrained planner sends a notification before authorization; examining that other path does not determine whether labeled queries and expected evidence distinguish retrieval from synthesis, which is what distinguishes the issue in the stated use case.

B:

The proposed probe concerns retraining uses revised labels and a different source snapshot, where dataset lineage and model versioning support repeatable assessment; this incident instead requires evidence about labeled queries and expected evidence distinguish retrieval from synthesis and the failure that generator tuning fails because the required documents were never retrieved.

D:

Inspecting transcript alignment and transcription errors explores a separate connection, where speech transcription transforms audio to text for subsequent analysis; the diagnosed requirement is stakeholders report hallucinations but cannot isolate search errors, and the actual defect is generator tuning fails because the required documents were never retrieved.

Question 825

A product owner must measure whether an AI-enabled change actually improves this process: Teams want approved external records searchable in Copilot. The earlier problem was: The agent lacks knowledge of non-Microsoft content. Which outcome indicator is most relevant?

  1. Differentiated requirements satisfied without unsupported workarounds
  2. Task completion without switching application context
  3. Relevant external content found without access violations
  4. Cross-domain tasks routed to the correct specialist

Explanation

Correct Answer

C:

The observed shortcoming is that the agent lacks knowledge of non-Microsoft content; measuring relevant external content found without access violations gives business owners a directly relevant indicator for the process where teams want approved external records searchable in Copilot.

Incorrect Answers

A:

Differentiated requirements satisfied without unsupported workarounds; that indicator describes a distinct workflow where competitive requirements depend on unique multi-model control; it does not quantify relevant external content found without access violations, which is needed for the stated case.

B:

The proposed metric concerns users must remain in a proprietary work-order application, whereas the failure in this case is that the agent lacks knowledge of non-Microsoft content; the indicator needed here is relevant external content found without access violations.

D:

For the workflow where a request spans procurement, finance and policy expertise, cross-domain tasks routed to the correct specialist is an actionable measure; it does not establish whether the agent lacks knowledge of non-Microsoft content in the situation under review.

Question 826

A staging test has to detect the specific defect that router saves tokens but sends difficult tasks to weak models. In the workflow where real task mix differs from prelaunch performance tests, which exercise verifies the corrective behavior?

  1. Query immediately before and after an effective-date change
  2. Re-evaluate the router on production-like stratified examples
  3. Seed lookalike records and check correct entity separation
  4. Prototype a unique scheduling dependency with the custom runtime

Explanation

Correct Answer

B:

The historical failure is that router saves tokens but sends difficult tasks to weak models; for the setting where real task mix differs from prelaunch performance tests, re-evaluating the router on production-like stratified examples verifies the dependency where feedback-driven thresholds rebalance accuracy, latency and spend.

Incorrect Answers

A:

For policy documents contain draft, superseded, and currently effective revisions, the alternate test is about whether version and effective-date metadata select authoritative content at retrieval time; it cannot prove the outcome needed after the reported regression where router saves tokens but sends difficult tasks to weak models.

C:

The tested behavior is that canonical keys and deduplication prevent incorrect cross-record joins; the necessary evidence here concerns feedback-driven thresholds rebalance accuracy, latency and spend and the previously observed router saves tokens but sends difficult tasks to weak models.

D:

This check exercises a different system behavior: owning the engine provides control over models, routing and execution; a pass there cannot establish that router saves tokens but sends difficult tasks to weak models no longer occurs in the required workflow.

Question 827

A planned AI solution must support this situation: The same agent is offered in voice and asynchronous chat. Its prototype fails because voice responses contain unreadable links and lengthy tables. Which design best fits the constraint?

  1. Use agent reasoning for uncertain multi-step analysis with bounded tools
  2. Use channel-specific accessibility and response policies
  3. Add approved case-summary guidance to the service workflow
  4. Design graceful agent-to-agent and human escalation on low confidence

Explanation

Correct Answer

B:

Because channel-aware response design adapts wording and modality, this approach corrects the reported breakdown where voice responses contain unreadable links and lengthy tables; the affected situation is that the same agent is offered in voice and asynchronous chat.

Incorrect Answers

A:

The alternative addresses a simple scripted topic cannot handle legitimate variants by a design in which reasoning can plan through evidence and tools within enforced boundaries; it does not supply the missing relationship for the same agent is offered in voice and asynchronous chat, where voice responses contain unreadable links and lengthy tables.

C:

Summaries omit customer impact and current ownership; the alternate design can address that situation because tailored instructions and contextual fields standardize case handoff content, whereas this question needs a solution to voice responses contain unreadable links and lengthy tables.

D:

For another type of architecture, confidence, retry ceilings and escalation routing enable resolution is useful; this design must instead support the same agent is offered in voice and asynchronous chat despite the problem where voice responses contain unreadable links and lengthy tables.

Question 828

Governance approves the business purpose but requires control of this exposure: Audit logs might be altered to obscure who deployed a model. The agent will be used when a model is replaced without a traceable change ticket. Which boundary is most defensible?

  1. Prevent mixed-version retrieval without compatibility checks
  2. Reject mismatched schema/model combinations
  3. Require reviewed datasets and independent promotion sign-off
  4. Require new data rights and labeling checks before training
  5. Restrict modifications and protect audit storage from tampering

Explanation

Correct Answer

E:

Avoiding the consequence where audit logs might be altered to obscure who deployed a model requires a control for a model is replaced without a traceable change ticket; in this architecture, deployment audit links actor, approved artifact and release time, which is the point at which the boundary can be applied.

Incorrect Answers

A:

For the use case where a search index changes embedding model and chunking approach, the control addresses the threat that mixed embedding or index versions might corrupt retrieval quality; the specified agent must be protected against audit logs might be altered to obscure who deployed a model.

B:

For the use case where a retraining release changes the tokenizer and label order, the control addresses the threat that a new tokenizer might be deployed with an old label mapping; the specified agent must be protected against audit logs might be altered to obscure who deployed a model.

C:

Business analysts can submit examples but not change production behavior; its guardrail addresses the possibility that unreviewed training examples might poison a promoted model; the present architecture must constrain the different threat that audit logs might be altered to obscure who deployed a model.

D:

Where customer terminology changes over time, the guard can address retraining might incorporate sensitive examples without renewed data rights review; the exposure that audit logs might be altered to obscure who deployed a model remains relevant to the present assessment.

Question 829

Risk reviewers need a control, not merely a monitoring dashboard. The use case is: Customer-facing brands have several names for the same product. Potential harm: An ambiguous product nickname might map to an unrelated product line. Which protection fits?

  1. Do not infer product identity when aliases collide
  2. Reject or escalate uncertain cross-system matches
  3. Validate source access after ranking and before response
  4. Enforce source-system entitlements in the integration

Explanation

Correct Answer

A:

Customer-facing brands have several names for the same product; its risk scenario is that an ambiguous product nickname might map to an unrelated product line; the safeguard fits because curated alias-to-product mapping grounds interpretations in the product catalogue and the restriction addresses that specific exposure.

Incorrect Answers

B:

The alternate guard addresses a guessed cross-system account match might attach credit risk to the wrong customer for two systems share company names but not primary keys; the present threat, an ambiguous product nickname might map to an unrelated product line, needs a different enforcement boundary.

C:

The proposed control targets a reranker might return a relevant passage that the user cannot access, while the given risk scenario is an ambiguous product nickname might map to an unrelated product line; these exposures arise in different operational contexts.

D:

That protection is tied to the mechanism that a supported connector can retrieve external information with explicit identity mapping; it does not prevent the distinct risk where an ambiguous product nickname might map to an unrelated product line in the system under review.

Question 830

Security and business owners want to keep this capability: Deployment automation pulls model artifacts from unverified origins. They must also prevent the following: An attacker might swap a model artifact in the deployment pipeline. Which guardrail should the design incorporate?

  1. Keep enough prior approved artifacts for recovery
  2. Restrict production to trusted vetted model versions
  3. Restrict modifications and protect audit storage from tampering
  4. Reject unregistered or unapproved models from production

Explanation

Correct Answer

B:

Avoiding the consequence where an attacker might swap a model artifact in the deployment pipeline requires a control for deployment automation pulls model artifacts from unverified origins; in this architecture, artifact provenance and integrity checks guard model supply chain, which is the point at which the boundary can be applied.

Incorrect Answers

A:

The alternate guard addresses a rollback might restore a model but not its required preprocessing artifacts for a preprocessing release changes category labels incorrectly; the present threat, an attacker might swap a model artifact in the deployment pipeline, needs a different enforcement boundary.

C:

The alternate boundary mitigates audit logs might be altered to obscure who deployed a model, which is relevant to a model is replaced without a traceable change ticket; it leaves the current workflow vulnerable to an attacker might swap a model artifact in the deployment pipeline.

D:

This restriction is associated with data scientists cannot explain why a newer classifier behaves differently and addresses the exposure that a model outside the approved registry might reach production; it does not specifically protect the target workflow against an attacker might swap a model artifact in the deployment pipeline.

Agentic AI Business Solutions Architect (AB-100) Practice Tests